Niko1221 cc3f9142e3 Auto on the 12 GB / streaming CUDA class: IQ3_S starts the MTP, UD-IQ4_XS-v31 starts DFlash2 with the tuned pack at the class's own 0.8 ...

drafting.CLASS_DEFAULTS: one removable entry per set, read from the signals the confidence already reads (a CUDA card whose engine slots are under the last-but-one CONF_STEPS step).
IQ3_S: on P2003 the file cache makes the MTP faster than DFlash2 by 12-44% (RTX2 t4b). UD-IQ4_XS-v31: DFlash2 beats the MTP by about 6.5% (t8, 3 starts; DF-VAL forced-fast tuned 0.8 x1.037).
The v31 entry DEPENDS ON RTX2's FILE-TIER FIX (unbuffered reads) BEING IN THE SAME RELEASE; without it remove that entry.
Anything chosen by name wins. Resident CUDA, Strix, gfx1201, Arc and Metal entries and the conf tables are unchanged. The sm_86 support line reads the set: UD-IQ4_XS-v31 gets the existing
'faster on some kinds of text and slower on others' line, every other model keeps 'slower'.

(cherry picked from commit 42412d42ff835d6a66a78fbf6d119911908eda1b)

2026-11-06 15:47:42 +01:00

..

__init__.py

terminal: strata_cli, strata_control, bin/strata and the terminal tests

2026-10-15 10:23:23 +02:00

acquisition.py

start and serve never acquire: one guard, a typed error, a test that fails on egress

2026-10-15 11:52:21 +02:00

amd_four_card.py

Four-card AMD serving profile: 32 slots in 4 groups, split 13,26,39, chunk 2048 (setup and launcher)

2026-10-29 15:06:10 +01:00

arg_layers.py

launch: --set mmap-experts takes a word for strata-dsv41 and stays a bare switch for Qwen (#5027)

2026-11-04 10:54:12 +01:00

atomic.py

Time-skip and tune job records are replaced with a wait for a reader (Windows 'Access is denied' when the host reads the file at that moment: it ended jobs at random)

2026-10-16 13:29:12 +02:00

autotune.py

Merge f-0213-life (final 576d06553) into f-0213-int

2026-11-04 12:24:01 +01:00

benchmarking.py

POSIX owner-only checks: repair our own folders whose mode drifted, note drives without modes, and name the chown command for folders another account owns

2026-10-20 17:00:25 +02:00

calibration_jobs.py

POSIX owner-only checks: repair our own folders whose mode drifted, note drives without modes, and name the chown command for folders another account owns

2026-10-20 17:00:25 +02:00

calibration.py

launch: one precedence for engine options; an explicit --set draft-conf and calibrated values survive (#5178, #5132, #4985)

2026-11-04 10:36:06 +01:00

cfg_paths.py

A saved run config follows the update: paths into a removed Control, engine or DeepSeek version folder are pointed at the current one (#4891)

2026-11-02 08:08:13 +01:00

components.py

tls: downloads check the server against the operating system's certificates; a failed check gives a plain message (#3722)

2026-10-22 15:43:12 +02:00

default_context.py

128K on every card, first-run setup and DeepSeek included: one rule (strata_control/default_context.py) for the app, the Terminal launch default, setup and the MCP install default; the tight-card fallback and DeepSeek's fixed 8K are gone, the inline warning stays

2026-11-01 17:35:17 +01:00

discovery.py

launcher: draft layer also found in --data-dir, split engine counts free VRAM, plan card numbers are the --gpu numbers

2026-11-04 07:11:31 +01:00

doctor.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

draft_layer.py

Models: a per-model draft layer (MTP) switch, picture-off and custom context that stick

2026-10-19 19:29:39 +02:00

drafting_fetch.py

DFlash2 on Intel Arc: the tuned bf16-selector pack for SYCL + IQ3_XXS at Arc's own confidence 0.6

2026-11-06 15:47:42 +01:00

drafting.py

Auto on the 12 GB / streaming CUDA class: IQ3_S starts the MTP, UD-IQ4_XS-v31 starts DFlash2 with the tuned pack at the class's own 0.8

2026-11-06 15:47:42 +01:00

driver_floor.py

CUDA 12 driver floor: setup takes it from the engine package, not from a constant (#4247, #4188)

2026-10-28 23:53:00 +01:00

dsv41_context.py

DeepSeek-V4.1-Flash: 512K and 1M contexts in setup and the Terminal, as Project Maya v1.1.28 did (#442 by @needmorevram, MIT)

2026-10-26 13:25:32 +01:00

dsv41_engine.py

DeepSeek's pictures on a Mac follow the installed engine: on exactly when the strata-dsv41 program carries STRATA_DS4_METAL_VISION_V1 (read, never started, cached per path; no engine yet: on, and setup looks again once it is fetched). Setup settles it after the engine step (no marker: the picture file is not fetched, pack and engine text only, with the plain line), the Mac model rows and the install tool ask the same function, the Coder's Mac line and docs/DEEPSEEK_MAC.md follow. The static DSV41_MAC_VISION is gone; other platforms are unchanged

2026-11-02 19:05:35 +01:00

dsv41_fit.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

dsv41_flags.py

DeepSeek --parallel: the engine flag probe lives in its own module (strata_control/dsv41_flags.py), because serving must not import dsv41_engine.py (it reaches the network)

2026-10-31 20:13:11 +01:00

dsv41_parallel.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

dsv41_paths.py

Engine update beside a loaded model: never by itself, the person's yes first, stop cleanly, start again with the same settings; a quiet notice and a once-a-day check in the launch monitor

2026-11-01 13:09:21 +01:00

dsv41_sets.py

Add the ISTA-DAS Coder set (192 experts a layer): one JSON for the repository, setup recipe, fit numbers, menu rows, Mac note, tests

2026-11-01 19:18:48 +01:00

engine_component.py

nvidia-smi is asked once per field and 30 s: one launch --plan on Windows with an engine installed started it 19 times

2026-11-04 20:24:56 +01:00

engine_library.py

Settings > Engine lists the installed engines (switch, delete, sizes); About only updates. Updates keep every older engine; strata components engines/use/remove share the app's code; DeepSeek has its own choice that launch, setup and the app follow

2026-11-01 12:01:33 +01:00

engine_update.py

After a model is stopped the update waits for its engine to let go; the start-again question follows what was changed

2026-11-01 17:15:20 +01:00

feed.py

feed: a drafters member for DFlash2 drafter files Strata hosts (pack-drafter, build-feed --drafters, make_local_feed --drafter)

2026-11-06 07:52:31 +01:00

folders.py

setup: banner names the chosen family; the user's own home (root's included) is allowed as a data folder; a missing DeepSeek engine package is said plainly (supported cards), the developer hint only with --verbose; tests

2026-10-24 00:05:11 +02:00

gguf_preflight.py

Own model files: the checker in front of setup accepts what setup --gguf-dir accepted (Unsloth split sets with a metadata-only first shard, links, several sizes in a folder, a folder above the model folder, more than 128 files); Unsloth Q5/Q6/Q8 and Swift IQ3_S are in the table

2026-10-20 17:00:01 +02:00

gpu_fit.py

Version 0.2.10 / engine 0.1.56 follow-through: test_tui2_023_versions and AGENTS.md follow the bump; the '0.2.9.1' tags that meant this release now read 0.2.10; the test comment names 0.2.9.1 as the unpublished DeepSeek hotfix label

2026-11-02 14:54:37 +01:00

gpu_pick.py

AMD cards fix, review round: setup's pinned one-card pick is still written (with gpus_by), only DeepSeek on AMD is judged when a copy adopts an earlier card, the launcher uses the plan's card rule and remembers a named card, Windows uses setup's amd_rank and no Linux arch list, message names strata launch --gpu

2026-10-31 03:05:41 +01:00

gpu_roles.py

owner decisions: plan cards keep the screens' count and add "(--gpu N)"; refusal says "free"; app tells the plan which engine it will unload

2026-11-04 09:32:34 +01:00

hipblaslt_tuning.py

hipBLASLt tuning table on by default on Strix Halo (gfx1151) only, the owner's choice; opt-in elsewhere

2026-10-19 21:37:42 +02:00

intel_spec.py

Setup and the launcher write --spec 2 for exactly two Arc B-series cards

2026-10-29 15:52:37 +01:00

journal.py

control: the operation record takes a long start wait too, and the tests walk the whole CLI path (#3459 #4191)

2026-10-28 23:33:57 +01:00

longpath.py

Engine install: survive Windows MAX_PATH (short staging folder, extended-length paths, plain file errors, deep-folder warning)

2026-10-20 11:15:20 +02:00

loopback.py

Calls to this computer's own servers skip the system proxy (tuner, tune session, launch self-check)

2026-10-20 17:00:00 +02:00

mac_fit.py

Mac app: Unsloth's sets are offered and installed as the Terminal's setup_mac does (metal_repack to one GGUF, --ple-gguf the table's shard, disk counts the repacked file); the repack tools ship in the installer bundle; tests

2026-11-05 09:05:01 +01:00

management.py

GUI: ask the installed Control for the start wait it actually takes, and let the CLI take the runtime's ceiling (#3459 #4191)

2026-10-28 23:33:57 +01:00

models_move.py

Models move: remember an unfinished move in the old folder; strata models move --resume/--leave/--status; one-line notice at strata launch

2026-10-20 16:53:14 +02:00

models.py

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00

offline.py

terminal: strata_cli, strata_control, bin/strata and the terminal tests

2026-10-15 10:23:23 +02:00

owner_transport.py

control: owner-only files on NTFS (a one-entry protected ACL via ctypes) and the owner state-file check uses it on Windows

2026-10-15 11:08:26 +02:00

paths.py

host: no /usr/bin/tar, /usr/bin/env or HOME/.local/state on Windows

2026-10-15 12:31:49 +02:00

pictures.py

Load with Read pictures on and no picture files: fetch only those files, keep the model's pack, draft layer and data folder, no installer run

2026-11-01 21:29:12 +01:00

posix_owner.py

POSIX owner-only checks: repair our own folders whose mode drifted, note drives without modes, and name the chown command for folders another account owns

2026-10-20 17:00:25 +02:00

preset_marks.py

Windows file in use: every state-file swap goes through atomic.replace (retry); a sharing violation no longer says run as yourself

2026-10-20 17:00:00 +02:00

presets.py

Presets: a value outside the range the tuner tries is kept with a warning, not dropped (#3844)

2026-10-27 02:02:46 +01:00

private_files.py

Windows: the token default DACL grants generic rights, so a child process can still be opened after adopt_user_defaults (#3428, #3399)

2026-10-26 18:21:07 +01:00

procs.py

serve: an engine left behind by an earlier run is waited for until it has given its memory back (#4051)

2026-10-29 00:01:54 +01:00

profiles.py

A saved run config follows the update: paths into a removed Control, engine or DeepSeek version folder are pointed at the current one (#4891)

2026-11-02 08:08:13 +01:00

progress_words.py

Progress words: a file being repaired after a wrong SHA-256 says so in plain words

2026-10-20 23:55:59 +02:00

readiness.py

API key is optional beyond this computer: warn, never refuse

2026-10-20 16:59:59 +02:00

recipes.py

POSIX owner-only checks: repair our own folders whose mode drifted, note drives without modes, and name the chown command for folders another account owns

2026-10-20 17:00:25 +02:00

replica_profile.py

Version 0.2.10 / engine 0.1.56 follow-through: test_tui2_023_versions and AGENTS.md follow the bump; the '0.2.9.1' tags that meant this release now read 0.2.10; the test comment names 0.2.9.1 as the unpublished DeepSeek hotfix label

2026-11-02 14:54:37 +01:00

runtime.py

runtime owner: answers status while it plans and launches, gives the first token its own budget, leaves when its server is gone, follows a process it cannot read only after repeated looks, and is recoverable before its first launch (#5177 #5059 #4293 #4574 #4893 #4250 #4973 #4824)

2026-11-04 10:40:01 +01:00

signing.py

updates: signed feed, Ed25519 verifier, component updater with resume, staged test-start, switch and rollback

2026-10-15 17:13:41 +02:00

speed_switches.py

Residency-biased routing (STRATA_ROUTE_PRIOR) is off by default again on 12 GB NVIDIA cards (#4402)

2026-10-29 10:25:07 +01:00

split_engine.py

Run manifest (lite): STRATA_MANIFEST=<file> writes the selected fast paths, fallback reasons, tensor formats, engine choice and budgets as JSON (P0.1)

2026-11-06 00:21:53 +01:00

sysinfo.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

targets.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

time_skip.py

Control: the time-skip check and tune engine reads also tolerate non-UTF-8 output (#4675)

2026-10-30 19:12:42 +01:00

tls.py

tls: downloads check the server against the operating system's certificates; a failed check gives a plain message (#3722)

2026-10-22 15:43:12 +02:00

transfer.py

Feed reads: one more try when the first times out or the connection fails (the first wait is 12 s, the second the full 30 s); about one github.com request in 25 hung for 30 s and the first-run card said offline

2026-10-24 14:44:12 +02:00

tune_api.py

Calls to this computer's own servers skip the system proxy (tuner, tune session, launch self-check)

2026-10-20 17:00:00 +02:00

tune_apply.py

Windows file in use: every state-file swap goes through atomic.replace (retry); a sharing violation no longer says run as yourself

2026-10-20 17:00:00 +02:00

tune_engine.py

Control: the time-skip check and tune engine reads also tolerate non-UTF-8 output (#4675)

2026-10-30 19:12:42 +01:00

tune_history.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

tune_knobs.py

launch: --set mmap-experts takes a word for strata-dsv41 and stays a bare switch for Qwen (#5027)

2026-11-04 10:54:12 +01:00

tune_machine.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

tune_session.py

tuning: the worker ignores STRATA_API_KEY, restart candidates are judged against the base's own free memory, an engine abort shows its own error line and is an OOM (#5056, #5122, #5078)

2026-11-04 10:44:36 +01:00

tuning.py

Control: the OS and architecture are read without WMI on Windows, so a stuck winmgmt cannot hang models list or a load (#4975, #5185 as reference)

2026-11-04 10:26:40 +01:00

uma_pool.py

Launch: a remembered card choice is said when the PC has more cards (NVIDIA DeepSeek after --gpu 0 stayed on one card without a word); Strix Halo: a fixed small expert cache lowers setup's own RAM budget (the person's cache and budget are only warned about) and a silent exit code 1 gets a plain next step

2026-11-01 09:51:26 +01:00

update_keys.json

Release wiring: release feed key and Tauri updater key, GitHub Releases feed address in the installers, versions 0.1.45.3 and 0.2.0-beta

2026-10-19 07:23:44 +02:00

updater.py

update --check on a source checkout prints the tree's own version for Control instead of '?' (display only; decide_control is unchanged)

2026-11-04 17:20:23 +01:00

user_args.py

Fix #3504: a two-value option (control-vector-layer-range) is written and removed as two argv entries, and an empty Advanced value is no longer a bare flag

2026-10-21 00:02:01 +02:00

vision_reader.py

Docs: Qwen picture reader on demand (DETAILS vision table, vision_reader note)

2026-11-02 02:01:12 +01:00

vram_fit.py

Layer split spill walk: stage-weight trim for short cards, drafter drop and head on host per card, KV window and fewer-cards steps; launcher retry without a drafter and the engine's own reason as the last line

2026-11-03 10:02:18 +01:00

winproc.py

Fix #3488: a reused PID in a child marker without a start time no longer blocks every load; the engine hash check reads twice before refusing

2026-10-20 23:53:07 +02:00